by Laxima.See every AI on your computer.
Models, the programs running them, the servers your Wi-Fi can reach and the agents that act with your permissions. Quarantai puts all of it on one map.
01 · The file
A model lands on your Mac.
A few gigabytes of weights, saved by an app, a script or a download you’ve forgotten about. Quarantai finds it, even under another name.
02 · The program
A program loads it.
Now the model runs on your memory and your processor. Quarantai shows which program has it open.
03 · The network
Your Wi-Fi can reach it.
One setting, and every device at home can use the model too. Quarantai tells you who can reach it.
04 · The agent
A script starts using it.
It sends the model instructions and acts on the answers, with your permissions. Quarantai flags it, even when it has no name you’d recognize.
05 · The map
Quarantai shows you all of it.
Who uses AI, which model, where it runs and who can reach it, on one map. Each warning sits on the part it’s about.
Why antivirus misses it
There is no signature for a language model.
A model nobody approved looks just like one you downloaded on purpose: ordinary files, run by ordinary and often well-known software. Your antivirus has nothing to match it against. Quarantai looks at your Mac the way you would if you had a free week: what’s running, what it loaded, who can reach it and how it got there.
Found by what’s inside.
A model saved as holiday.jpg still shows up as a model.
Judged by what it does.
A home-made script gets the same attention as a famous app. A familiar name earns no free pass.
Open doors, spotted.
For every AI server you see who can reach it: only your Mac, everyone on your Wi-Fi, or any web page you open.
Quiet about your own models.
Approve the models you use on purpose, and Quarantai stops asking about them. If a file changes, it asks again.
Under the hood
What a scan covers
Here’s what Quarantai checks, in plain terms. We keep the exact rules to ourselves, because a published list of what we look for would double as a list of what to avoid.
- Model files
- Weights in the widely used formats, anywhere in your home folder. Files that would run code the moment they’re opened, or that are built to trip up the program loading them. Copies of the same model in several places, the trail a spreading model leaves behind.
- Code that comes with a model
- Scripts and settings a model runs when it loads, checked for things a model has no business doing.
- Programs running AI
- Which programs have a model loaded right now, including ones running from odd places or dressed up as something else.
- Who can reach it
- For each AI server: only this Mac, your home network, or web pages in your browser. And whether anyone has already connected from outside.
- AI agents
- Coding agents and home-made scripts that use AI and act on your Mac. Each session gets its own line, with how it was started and whether it was told to skip asking you first.
- Agent setup
- The tools and instructions your agents are configured with, including text hidden from you that a model can still read.
- What starts by itself
- AI set to start at login, on a schedule or with your terminal, so you know what comes back after a restart.
- Remote access
- Remote-login and screen-sharing services that other devices on your network can reach.
Quarantai only reads. It never runs a model and changes nothing on your disk. It treats every file as hostile, and it won’t make iCloud download files you keep only in the cloud.
Built for the Mac first. The command-line scanner also runs on Linux.
Coming next · Quarantai Guard
AI agents work with your permissions. Guard will watch what they do with them.
Every command, live. What each agent runs and where it connects.
Your rules. “Ask me before git push.” “Never run a script straight from the internet.” Guard holds anything that breaks them.
A check for prompt worms. Instructions that try to copy themselves from one agent to the next get held before they spread.
The scan is free. Guard, the real-time part, will be a paid subscription.
- 09:41:18Cursorreply repeats its own promptprompt-worm check
- 09:41:15Claude Codegit push origin mainyour rule: ask first
- 09:41:11agent.pyconnect 192.168.1.23:9931reaching a neighbor
- 09:41:09agent.pycurl -s https://… | shwaiting for you
- 09:41:04Claude CodeTLS api.anthropic.comAnthropic
- 09:41:02Claude Codenpm testran
Numbers from real runs
- files
- 3.18Mfilesscanned in 92 seconds: the whole home folder of a developer’s Mac
- false alarms
- 0false alarmson a developer’s everyday Mac. The first run raised 423 warnings, and we tuned on that data until it raised none.
- broken files
- 84,000broken filesfed to Quarantai on purpose. It read every one without a crash.
Protect your data from a new kind of threat.
Questions
What is Quarantai?
Quarantai is an app by Laxima that shows you every AI on your computer: model files, the programs running them, model servers other devices can reach, and AI agents that run commands with your permissions. It only reads and reports. It never runs a model and changes nothing on your Mac.
What counts as AI nobody approved?
Anything AI-related on your computer that you didn’t knowingly set up, or that does more than you agreed to. A model some program downloaded in the background. A model server your whole network can reach. An agent that runs commands without asking, or one started by a package you installed for something else.
Does Quarantai replace my antivirus?
Keep your antivirus. Quarantai covers what it can’t see. An antivirus looks for known malicious code, and a local model is ordinary files run by ordinary, often legitimate software, so there’s nothing for it to match. Quarantai looks at what’s actually on your Mac and what it’s doing.
Do my files or scan results leave my Mac?
No. The scan runs on your Mac, and your files and results stay there.
Will it flag the models I use on purpose?
Once. Approve your own models and Quarantai remembers them, so it stops asking. A model sitting on your disk is never an alarm by itself. Warnings come when something around it is wrong, like hidden code, a server open to your network or an agent nobody recognizes. On a developer’s everyday Mac the first run raised 423 warnings. After tuning on that data, it raised none.
Does it see Claude Code, Cursor and other coding agents?
Yes. Quarantai lists the AI agents working on your Mac, such as Claude Code, Cursor or Codex, with each session and how it was started. It also catches agents with no name at all, like a script someone left in Downloads. A well-known agent used normally is simply listed. Warnings come when one was told to skip asking you, was started by something you didn’t run, or runs from a place programs shouldn’t.
What does it cost?
The scan is free, and it will stay free. Guard, the real-time part coming next, will be a paid subscription. We’ll publish the price before Guard launches.
What is Quarantai Guard?
The next part of Quarantai, in development now. Guard will watch AI agents in real time: every command they run and every connection they open. You’ll be able to set rules, like “ask me before git push”, and Guard will hold anything that breaks them. It will also catch instructions that try to copy themselves from one agent to another, the way a prompt worm spreads. It starts in a mode that only shows what it would have stopped.
Does it run on Windows or Linux? How do I get it?
Quarantai is built for the Mac first: the Mac app with the map is in development, and the command-line scanner already runs on macOS and Linux. Windows isn’t supported yet. Join the waitlist on this page and we’ll email you when it’s your turn.
Names of other products and models on this page belong to their owners and are used only to identify the software Quarantai recognizes. Quarantai is not affiliated with or endorsed by them.